Session Slot: 10:30-12:20 Monday
Estimated Audience Size: 200
AudioVisual Request: One Overhead Projector
Session Title: Practical Power Analyses for Statistical Consulting
Theme Session: No
Applied Session: Yes
Session Organizer: Glueck, Deborah H.
Address: Room N-101A University of Medicine and Dentistry of New Jersey Robert Wood Johnson Medical School 675 Hoes Lane Piscataway, New Jersey 08854-5635
Session Timing: 110 minutes total (Sorry about format):
Opening Remarks by Chair - 5 minutes First Speaker - 25 minutes Second Speaker - 25 minutes Third Speaker - 25 minutes Fourth - 25 minutes Floor Discusion - 5 minutes
Session Chair: Muller, Keith E. University of North Carolina-Chapel Hill
Address: Department of Biostatistics McGavran-Greenberg, CB#7400 University of North Carolina Chapel Hill, NC 27599
1. Power and Sample Size for Clinical Trials of Survival
Shuster, Jonathan J., University of Florida
Address: Director, Pediatric Oncology Group Statistical Office University of Florida 104 N. Main St. #600 Gainesville, FL 32601
Abstract: This paper will be devoted to methodology for computing sample size requirements for randomized two-arm clinical trials, whose endpoints are times until an event, such as survival or event-free survival. The methodology presumes that patients enter according to a Poisson Process, and that proportional hazards can be assumed.
Three scenarios will be considered. First, we shall derive the sample size for the situation where no sequential monitoring is utilized. This is covered in the author's book, ``Practical Handbook of Sample Size Guidelines for Clinical Trials'' (CRC Press 1992). Two sequential designs will utilize an O'Brien-Fleming approach. We shall derive the required properties where continuous sequential monitoring is feasible, thanks to advances in information transfer. Finally, we shall consider a group sequential approach, which is the first to directly plan for outcome-based timing of looks on the information scale, from the onset of the trial. (Lan and DeMets have looked at the robustness of changing the frequency of looks during the trial, when you have a Z-value within 20% of the O'Brien-Fleming bound. See Biometrics 45, pp. 1017-1020, 1989). At each look where significance has not been attained, the analysis leads to calculation of the timing of the next look, according to the closeness to the stopping barrier. If the calculated timing is higher than 100% information, the trial is stopped immediately for non-significance.
For each of these sequential methods, the maximum amount of information over and above that of a non-sequential method is remarkably small. Major gains in terms of average information are realized. Maximum sample size requirements for either of the two sequential methods can be derived by multiplying the non-sequential sample size requirement by a relative efficiency factor that depends only upon the Type I and Type II errors.
2. Power in the General Linear Multivariate Model
Glueck, Deborah H., Robert Wood Johnson Medical School
Address: Room N-101A University of Medicine and Dentistry of New Jersey Robert Wood Johnson Medical School 675 Hoes Lane Piscataway, NJ 08854-5635
Abstract: The general linear multivariate model (GLMM) is used often in clinical and epidemiologic experiments. In contrast to data analysis, power calculations depend on the distributional properties of the predictors. The conditional power is calculated for a specific realization of the predictors. We describe a general class of conditional power approximations based on the method of moments and Taylor series. We demonstrate how existing conditional power approximations fall into this scheme. The unconditional power equals the expected value of the conditional power with respect to all possible stochastic realizations of any random predictors. The quantile power is the conditional power for a `worst case' scenario. We summarize unconditional and quantile power results for GLMMs with fixed and Gaussian predictors. An example power analysis illustrates the use of quantile power approximations in study design.
3. Power for Categorical Data Analysis
O'Brien, Ralph G., Cleveland Clinic Foundation
Address: Director, Collaborative Biostatistics Center Department of Biostatistics and Epidemiology Cleveland Clinic Foundation/P88 9500 Euclid Ave. Cleveland, OH 44195
Shieh, Gwowen, Cleveland Clinic Foundation
Abstract: From this broad topic we will focus on a pragmatic way to determine power for likelihood ratio tests in applications of logit, log-linear, and Poisson modeling from the simple to the complex. Specifically, we will show via examples how the dominating term of a good general noncentrality approximation for generalized linear models (Self, Mauritsen, and Ohara; Biometrics, 1992) can be computed by forming and analyzing ``exemplary'' data using one's regular statistical software (as per O'Brien; SUGI-11, 1986). Powers or minimal sample sizes are then computed, tabled, and/or graphed using UnifyPow, a freeware SAS macro (http://www.bio.ri.ccf.org/power.html). The handout will also cover how UnifyPow easily handles statistical planning problems for many other types of categorical data analyses.
4. Sample Size and Power Calculations for Studies with Correlated Data
Liu, Guanghan (Frank), Merck Research Laboratories
Address: Clinical Biostatistics Merck Research Laboratories 10 Sentry Parkway, BL3-2 Blue Bell, PA 19422
Abstract: Correlated data occur frequently in biomedical research and have received a good deal of attention in recent years. Examples include longitudinal studies, family studies, and opthalmologic studies, etc. Methods to analyze correlated data have been extensively studied in recent years. In this talk, we present methods to compute sample sizes and statistical powers for studies involving correlated observations (Liu and Liang; Biometrics, 1997). This is an important step when the investigators are planning a study to address some scientific hypotheses. The proposed method is based on generalized estimating equations (GEE). The theoretical background and some additional assumptions needed for the sample size and power calculations are highlighted. Sample size and power calculation formulas will be obtained for some special cases that are commonly seen in practice. The methods will be illustrated by simulations and examples from various clinical studies including longitudinal and cross-over trials.
List of speakers who are nonmembers: None